Information Types in Product Reviews

TL;DR

This work addresses identifying and leveraging information types within product reviews by introducing a broad typology of 24 sentence-level types and a zero-shot multi-label classifier built on FLAN-T5-xxl. It demonstrates that combinations of these types yield strong signals for downstream tasks such as review helpfulness ($F_1$ signals) and sentiment polarity, while revealing cohesive patterns in the rhetorical structure of reviews versus summaries. The authors validate their approach on the AmaSum dataset and show category- and text-level insights, including that summaries communicate core points more efficiently than full reviews. The proposed typology enables explainable analyses, targeted content filtering, and improved downstream modeling in the product-review domain, with potential extension to other domains.

Abstract

Information in text is communicated in a way that supports a goal for its reader. Product reviews, for example, contain opinions, tips, product descriptions, and many other types of information that provide both direct insights, as well as unexpected signals for downstream applications. We devise a typology of 24 communicative goals in sentences from the product review domain, and employ a zero-shot multi-label classifier that facilitates large-scale analyses of review data. In our experiments, we find that the combination of classes in the typology forecasts helpfulness and sentiment of reviews, while supplying explanations for these decisions. In addition, our typology enables analysis of review intent, effectiveness and rhetorical structure. Characterizing the types of information in reviews unlocks many opportunities for more effective consumption of this genre.

Figures (6)

• Figure 1: A sentence from a review, and the dominant information types communicated in it.
• Figure 2: Aggregated Type probability vectors over three classification tasks. Review-level probability vectors (\ref{['fig_classification_vectors_hr']} and \ref{['fig_classification_vectors_sentiment']}) are computed as the average of the sentence vectors contained in the review.
• Figure 3: Aggregated Type probability vectors in AmaSum brazinskas2021amasum reviews and summaries.
• Figure 4: Aggregated review-level probabilities of select Types, grouped by product categories.
• Figure 5: Expected probabilities of Types over the progression of a 6-sentence review in AmaSum.